Perplexity intuitively means we are trying to guess the number of neighbors around a point. This is why we keep it low for dense regions since there is a large cluster of points around each point.
Good you asked. In my video, I cover the gaussian formula and compare it back to back with the t-distribution. The literature calls the gaussian version of formulation as SNE algorithm. In practice, SNE's embeddings lag behind TSNE in providing clarity when you have large dimensions and some non-linearity. Have a look at the video's formulations for SNE's q version as it is compared to TSNE. Happy to answer should you have any other questions.
![Math for Game Devs [2022, part 1] • Numbers, Vectors & Dot Product](http://i.ytimg.com/vi/fjOdtSu4Lm4/mqdefault.jpg)
Why to keep perplexity value small(low) for dense region?
Perplexity intuitively means we are trying to guess the number of neighbors around a point. This is why we keep it low for dense regions since there is a large cluster of points around each point.
@@machinelearningmastery how do we know if regions are dense, a priori?
I have a question why not represent the low dimension probability ‘ q’ with gaussian rather than using t distribution
Good you asked. In my video, I cover the gaussian formula and compare it back to back with the t-distribution. The literature calls the gaussian version of formulation as SNE algorithm. In practice, SNE's embeddings lag behind TSNE in providing clarity when you have large dimensions and some non-linearity. Have a look at the video's formulations for SNE's q version as it is compared to TSNE. Happy to answer should you have any other questions.